By Kenrick Cai
SAN FRANCISCO, Aug 12 (Reuters) - Google co-founder Sergey Brin in recent months has urged key AI staff to go all in on the company's Gemini model as parent Alphabet seeks to close the gap with rivals, two people familiar with his remarks told Reuters.
As Anthropic raced ahead with a preview of its powerful Claude Mythos model, Brin addressed hundreds of employees in an April town hall, urging the company’s Google DeepMind AI lab to move faster, said one of the people on condition of anonymity.
Google’s Gemini model briefly overtook competitors last November, but new updates from Anthropic and OpenAI once again put it in catch-up mode. By August, Google had delayed the new flagship version of Gemini by two months after internal testing showed its performance continued to lag behind rivals in areas such as coding, one of the sources and another person familiar with the matter said.
On August 5, the Mountain View, California, tech giant announced a sweeping leadership overhaul of its Google DeepMind AI division in which its chief, Demis Hassabis, stepped aside in favor of his deputy, Koray Kavukcuoglu. Hassabis became chair, with fewer management obligations. At the same time, the two original technical co-leads of Gemini quit to co-found a startup.
While some industry observers have questioned if Google is exiting the model race, interviews with seven people knowledgeable about Gemini's development efforts show that Alphabet's commercialization strategy and reorganization of ranks last week are tied to its effort to achieve model supremacy.
The changes mark a further erosion of DeepMind's autonomy, which Google has chipped away at since buying the London-based lab in 2014 in an ongoing push for its AI to power more of its software and make more money. Alphabet’s shakeup included a revelation during an all-hands meeting on August 6 that some teams were moving out of DeepMind and into corporate Google, two sources said.
Internal meeting details and Brin's remarks have not previously been reported. Other media outlets earlier reported on not-yet-final reorganizations and Brin's coding push.
Google declined to answer questions from Reuters about Brin's recent involvement, DeepMind's structural changes, and concerns held by staff.
Under Hassabis, Google DeepMind pursued a wider range of AI research projects than its rivals while building Gemini models that are essential to corporate commercial interests. But in past years, Hassabis worked against some efforts that could have meant new revenue for Alphabet or helped it gain better footing in the AI race, a Reuters special report last year showed. His Google allies previously told Reuters that they rejected the idea that Hassabis gave short shrift to the business or focused too much on long-term science.
In 2025, Google CEO Sundar Pichai appointed Kavukcuoglu, then DeepMind’s chief technology officer, to an additional senior post as Google’s chief AI architect. His task was to embed Gemini across the company’s products. Four of the sources told Reuters that in the months that followed, some leaders at DeepMind found their influence diminished while Kavukcuoglu became the central figure shaping Gemini’s direction.
Kavukcuoglu had Brin’s backing, one of them said.
LEADERS SEEK TO CALM WORRIED STAFF
News of the shakeup caused Google’s stock to drop 4% the day of the announcement.
At the unit-wide all-hands meeting on Thursday of last week, Hassabis and Kavukcuoglu sought to assuage concerns by staff — most of whom learned about the change as it became public — about the AI lab's future.
The Google leaders said day-to-day work at the lab would remain the same, according to three sources with direct knowledge of the meeting. Kavukcuoglu would stay in charge of Gemini, the sources said; he had assumed oversight of model development more than a year prior, five sources said.
Hassabis gave a talk about how AI could advance science and alter society, long-term issues that would constitute the focus of his new role as chair, the three sources knowledgeable about the meeting said.
This messaging aside, the all-hands meeting surfaced fresh concerns that DeepMind’s autonomy was being subsumed by Google. Hassabis and Kavukcuoglu told staff that several nontechnical teams were being moved out of DeepMind and into the corporate reporting structure, two of the sources said. For an organization that had once encouraged employees to pursue ancillary projects, the message to reclaim model supremacy was also clear.
Executives including Pichai have said in recent months that Google considers itself to be at the frontier in certain areas like video generation, and remained committed to catching up in other areas despite setbacks.
Senior leaders in the cloud division, meanwhile, expect Kavukcuoglu’s appointment to alleviate tensions with DeepMind over allocation of Google’s constrained supply of AI chips, known as Tensor Processing Units (TPUs), another source said.
BRIN PUSHES RECURSIVE SELF-IMPROVEMENT
The Google shakeup comes as Brin has worked to fire up staff, something he also did after OpenAI’s ChatGPT launch in 2022 jolted Google.
Although he no longer holds an executive title, the 52-year-old Brin has been informally influencing model training efforts. For several months now, he has been remarking to key staff that they need to catch up to the frontier, according to one of the sources knowledgeable about his efforts.
Brin has used the implicit power he holds as Google’s co-founder to push resource allocation toward specific areas, such as recursive self-improvement, the point at which the technology can improve without human intervention, this source said.
Google declined to make Brin available for an interview.
Hassabis, meanwhile, decreased his involvement in managing Gemini and passed off that responsibility to his deputies shortly after originally getting the AI model off the ground in 2023, two of the sources said.
Though Google caught up to the frontier last November with the release of Gemini 3, it quickly fell behind again. Four sources characterized its laggard status as a result of lingering structural challenges that made it fall behind in the first place.
Constraints in computing capacity, for instance, and disagreements between Gemini’s many project leaders led to fewer resources for areas that would prove critical, such as coding, two of the sources said. Google’s bureaucratic nature also caused it to release models at a slower pace than nimbler rivals, the four sources said.
Google declined to comment on these claims, but Pichai told Wall Street analysts last month that the company would continue to scale up its infrastructure to alleviate constraints around TPUs that it needed for model training, as well as consumer and enterprise products.
POWER FURTHER SHIFTS TO CALIFORNIA
Against this backdrop, Kavukcuoglu has been consolidating power. He moved from London to Mountain View last year for the chief AI architect promotion, reporting to Pichai.
Some London-based leaders found their influence on Gemini's technical direction had diminished in the wake of Hassabis decreasing his involvement in Gemini and Kavukcuoglu moving to Google's headquarters, according to a source with direct knowledge.
Kavukcuoglu began building a reputation within other parts of Google as well, as someone more attuned to commercialization. Even before his relocation, Kavukcuoglu rather than Hassabis was the primary point of contact between DeepMind and Google Cloud, the company’s main AI revenue engine, according to a source with direct knowledge.
He now will have final say on major decisions for DeepMind, a Google spokesperson told Reuters last week.
Some of the highest-profile leaders other than Kavukcuoglu, including Gemini's original technical co-leads Jeff Dean and Oriol Vinyals, left as part of last week’s shakeup. Staff reporting to some of these leaders had held a rare degree of agency to conduct research not directly related to training Gemini, three sources said.
In the wake of their departures, Google did not immediately share plans for the fate of those teams or those projects, making some anxious that the lab’s bulwark against commercial interests would further erode.
(Reporting by Kenrick Cai in San Francisco; Editing by Kenneth Li, Jeffrey Dastin and Matthew Lewis)
